Yipeng Yu (俞一鹏)

This paper provides a deep research of deep research. We position LLMs and Stable Diffusion as the twin pillars of generative AI, and lay out a roadmap evolving from the Transformer to agents. We examine the progress of AI4S across various disciplines. We identify the predominant paradigms of human-AI interaction and prevailing system architectures, and discuss the major challenges and fundamental research issues that remain. AI supports scientific innovation, and science also can contribute to AI growth (Science for AI, S4AI).

Is the user’s current query input exactly what they intend to search for?" Our work aims to answer this question by determining, at each typing, whether the current query is complete. If so, a search is implicitly triggered in advance without waiting for user confirmation. This approach reduces response time and enhances the user search experience. Specifically, we propose TaoType, a client-side framework that introduces innovation in data sampling, feature selection, model design and training, and online strategy. Experiments in a leading mobile shopping application named Taobao validate its effectiveness, achieving offline precision/recall/accuracy of 0.7936/0.8196/0.7742, respectively, and decreasing online response time by 640.51±93.65 milliseconds, which is of great benefit to the search system.

In this paper, we propose RecGPT-Mobile, a framework that designs a lightweight LLM-based intent understanding agent to improve recommendation quality in mobile e-commerce scenarios. By deploying LLM directly on mobile devices, our approach can capture the evolving interests of users more quickly and adjust the recommendation results in real time. Extensive offline analyzes and online experiments demonstrate that our method significantly improves the accuracy of recommendation results, laying a practical path for LLM deployment in production-scale recommendation systems on mobile devices, as well as a scalable solution for integrating LLMs into real-world next-query prediction systems.

We introduce 1D-Bench, a benchmark grounded in real e-commerce workflows, where each instance provides a reference rendering and an exported intermediate representation that may contain extraction errors. 1D is short for one day, representing the efficient completion of design-to-code tasks in less than one day. Models take both as input, using the intermediate representation as structural cues while being evaluated against the reference rendering, which tests robustness to intermediate representation defects rather than literal adherence. 1D-Bench requires generating an executable React codebase under a fixed toolchain with an explicit component hierarchy, and defines a multi-round setting in which models iteratively apply component-level edits using execution feedback. Experiments on commercial and open-weight multimodal models show that iterative editing generally improves final performance by increasing rendering success and often improving visual similarity. We further conduct a pilot study on post-training with synthetic repair trajectories and reinforcement learning based editing, and observe limited and unstable gains that may stem from sparse terminal rewards and high-variance file-level updates.

In this work, we extend KGED to a four-class classification task to identify which element is incorrect or whether the triple is correct. Directly adapting existing LLM-based methods to this setting yields limited performance, as they fail to effectively exploit the topological information of KGs. To address this issue, we propose a novel topology-injected LLM framework, which precomputes high-order common neighbors between head and tail entities as topological evidence, thereby reducing complexity. Furthermore, we introduce a mixture-of-experts adapter that maps structural and topological evidence into the text embedding space.

We propose SEAL, a novel contrastive learning framework. It leverages structure-aware learning to preserve semantic hierarchies and masked element alignment for fine-grained semantic discrimination. Furthermore, we release StructDocRetrieval, a long structured document retrieval dataset with rich structural annotations. Extensive experiments on both the released and industrial datasets across various modern PLMs, and online A/B testing demonstrate consistent improvements, boosting NDCG@10 from 73.96% to 77.84% on BGE-M3.

This study aims to address this gap by incorporating users' emotional and behavioral indicators into the prediction of short-video viewing behavior. Study 1 was conducted in a controlled laboratory setting, where participants viewed videos on a computer screen while their physiological activity was recorded as an objective measure of emotional responses. Study 2 aimed to enhance ecological validity by having participants view videos on mobile phones, enabling them to swipe between videos as they would in a typical short-video app. Our findings provide valuable insights into the psychological drivers of short video viewing behavior and present a novel, non-intrusive approach to incorporate users’ real-time experiences into recommendation systems.

In this work we investigate the relevance of different positions of the latent image features to the target object in diffusion model, and accordingly propose a weighted-merge method to merge multiple reference image features into the corresponding objects. Next, we integrate this weighted-merge method into existing pre-trained models and continue to train the model on a multi-object dataset constructed from the open-sourced SA-1B dataset.

To free human from laborious video production, this paper proposes the building of VideoMaster, a multimodal system equipped with four capabilities: highlight extraction, video describing, video dubbing and video editing. It extracts interesting episodes from long game videos, generates subtitles for each episode, reads the subtitles through synthesized speech, and finally re-creates a better short video through video editing. To the best of our knowledge, VideoMaster is the first multimedia system that can automatically produce product-level micro-videos without heavy human annotation.

We present Text2Video, a novel system to automatically produce videos using only text-editing for novice users. Given an input text script, the director-like system can generate game-related engaging videos which illustrate the given narrative, provide diverse multi-modal content, and follow video editing guidelines. The system involves five modules.

In this paper, we build rat cyborgs to demonstrate how they can expedite the maze escape task with integration of machine intelligence. We compare the performance of maze solving by computer, by individual rats, and by computer-aided rats (i.e. rat cyborgs). They were asked to find their way from a constant entrance to a constant exit in fourteen diverse mazes. Performance of maze solving was measured by steps, coverage rates, and time spent. The experimental results with six rats and their intelligence-augmented rat cyborgs show that rat cyborgs have the best performance in escaping from mazes. These results provide a proof-of-principle demonstration for cyborg intelligence. In addition, our novel cyborg intelligent system (rat cyborg) has great potential in various applications, such as search and rescue in complex terrains.